[bvolpato]

Motivation

The GLM-5.3-Flash page says TileLang DSA with FP8 KV is not a valid CUDA combination. That is out of date: since #42957, the server accepts the pair when both backends are tilelang, so a reader can reach it with flags. On ba1ed00dfe, the pair fails under concurrent requests.

Modifications

One sentence now says that the server accepts the pair, that it is not qualified for this model, and that builds without #47013 can return a run of !. The deployment panel keeps its reason ("qualified on ROCm, not CUDA"), which is still correct.

#44217 also rewrites this sentence, so the second PR to land must resolve the conflict.

Accuracy Tests

This PR changes docs only. I served the real FP8 weights on 4x B200 and ran sgl-eval 0.1.3:

python3 -m sglang.launch_server --model-path <GLM-5.3-Flash FP8 weights> --served-model-name glm \
  --trust-remote-code --host 127.0.0.1 --port <port> \
  --tp-size 4 \
  --dsa-prefill-backend tilelang --dsa-decode-backend tilelang --kv-cache-dtype fp8_e4m3 \
  --moe-runner-backend flashinfer_trtllm \
  --speculative-algorithm EAGLE --speculative-num-steps 5 \
  --speculative-eagle-topk 1 --speculative-num-draft-tokens 6 \
  --reasoning-parser auto --tool-call-parser auto
sgl-eval run gsm8k --base-url http://127.0.0.1:<port>/v1 --model glm --num-examples 200 --num-threads 64 \
  --max-tokens 32768 --temperature 1.0 --top-p 0.95 --thinking

Both table rows differ from the cookbook's B200 Low Latency flags by --tp-size 4 instead of 8, --trust-remote-code, and --served-model-name glm. Their server_args differ only in the DSA backends, the port, and the random seed.

KV + DSA	Commit	GSM8K	Answers cut off at 32768 tokens
FP8 + TileLang	ba1ed00dfe	81.50%	16.5%
FP8 + TRT-LLM	582168dd7c	98.00%	0%

Each cut-off answer starts normally and then turns into a run of !. About 15 minutes into the eval, new short prompts returned ! from the first token with null logprobs. Prompts of 35k to 142k tokens got correct answers, but they ran one at a time after the eval, each after /flush_cache. This run does not show how long prompts behave under load, or whether short prompts recover. #47013 reports the same output on H200.

To test the link to #47013, I served 5-layer dummy weights at 825f8f6b89 with the same flags. At 33 and 41 concurrent requests, every request returned null logprobs and a run of !. At 19 or fewer, none did. With the #47013 kernel or TRT-LLM DSA, none did at 19, 33, 41, or 64. The #47013 regression test fails on the main kernel and passes with the #47013 kernel.

Not validated: real weights with #47013, real weights on other CUDA GPUs, and a docs site build. check_cookbook_configs.mjs and pre-commit ran on a trial merge with main.

Speed Tests and Profiling

Not applicable.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

---

CI States

Latest PR Test (Base): ✅ Run #41663412169
Latest PR Test (Extra): ❌ Run #41663411939
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.

Tip

Base CI only covers the core cases, and its workload may not exercise your change. Use /rerun-test to run the tests your change touches, or copy this prompt for an agent:

Find the registered tests this PR affects and run them with /rerun-test, following https://github.com/sgl-project/sglang/blob/main/.agents/skills/rerun-test/SKILL.md

[mintlify]

Preview approval required

This pull request comes from a fork, so Mintlify won't build a preview until someone with write access approves commit da27037.

Approve preview

[bvolpato]

Folded into #41898, whose error message this docs line contradicted.